AI.FO PROOF

Proven every night

Every night the whole engine is put through a synthetic cohort engineered so that every signal has somewhere to fire. The run publishes nothing unless the suite is green and every signal fires where it is meant to. The figures below are not a dashboard reading; each one is derived from artifacts committed to the repository, so anyone with the code can recompute them.

ARCHETYPE COVERAGE

100%48 of 48 signals in the current engine registry are mapped to a deterministic archetype engineered to force them to fire.

Coverage is a standing guarantee, not a one-off count. The mapping is pinned to the registry at build time, so a newly shipped signal cannot merge without its own archetype, and the nightly coverage gate fails the entire run if any mapped signal fails to fire. A published night is therefore a night on which the whole catalog was exercised, not merely a night the suite happened to pass. Every registered signal currently has a mapped archetype.

The cohort spans named archetypes, each a company built to make a group of related signals applicable and fire:

Archetype Liquidity SpiralArchetype Collections CrunchArchetype Margin SqueezeArchetype Payroll TreadmillArchetype Spend SprawlArchetype Forecast FogArchetype Rule of 40 ShortfallArchetype Net Revenue RetentionArchetype Debt Service StrainArchetype EBITDA ErosionArchetype Tax TrapArchetype Deferred Revenue BurnArchetype Owner Pay Sustainability

THE NIGHTLY LEDGER

These are the three figures the public telemetry artifact exposes, shown with the exact definition the artifact carries for each. They answer different questions and are deliberately kept apart rather than collapsed into one impressive-looking total.

14,070

Active assertions

The assertions the nightly suite actually executed tonight. This is the answer to "what is being checked right now".

45,035

Lifetime authored assertions

Every assertion ever promoted into a committed snapshot, archived and superseded runs included. This is the answer to "how much assertion evidence has been built". Until assertions have been archived or superseded, every authored assertion is also an active assertion, so this figure can match the active count exactly. That is a young archive rather than the same value published under separate labels. The figures separate once retention ages snapshots out, and also when an accepted baseline supersedes a generation, which leaves those assertions authored but no longer active.

365,098

Cumulative verifications

Verification work performed across every recorded run. This is the answer to "how much checking has actually happened", which is a different question from how many distinct assertions exist. The per-run ledger began empty, so early in its life this figure can match the active count exactly. That is a young ledger rather than the same value published under separate labels. The figures separate as CREDITED runs accumulate, which is not the same as runs accumulating: a run that contributed nothing still records an entry, so the match can persist while the ledger grows.

Most recent published run: 2026-08-26Commit 8c49f29aRun 2026-08-26-002

WHAT A PUBLISHED NIGHT DOES AND DOES NOT PROVE

What it proves

The suite was green at the stamped commit. A red suite publishes nothing at all rather than publishing a lower number, so an unchanged timestamp means a night did not promote, never that a night quietly passed.
Every signal in the registry fired in the archetype member mapped to it. Deterministic archetypes cover the full registry, and a signal that fails to fire fails the run and blocks promotion, so a published night is a night on which the whole catalog was exercised.
Every figure above was derived from artifacts committed to the repository at the stamped commit, not from a dashboard, a database, or a maintained spreadsheet. Anyone with the repository can recompute them.

What it does not prove

That the engine's financial judgments are correct. These figures measure that the engine is stable, internally consistent, and exercised across its whole catalog. Whether its conclusions are the right conclusions is a separate question, answered by expert review rather than by counting assertions.
That the engine works on real customer books. The nightly cohort is synthetic by construction, generated to span a wide revenue range and to exercise every signal. Synthetic data can be made adversarial, but it cannot surprise the way a real ledger does.
That the tests are independently audited. These are our own assertions, run on our own infrastructure. They are reproducible by anyone who clones the repository, which is a real and checkable claim, and it is a weaker claim than external verification.

VERIFY IT YOURSELF

The same figures, stamped with their commit and generation time, are public and need no login at /api/telemetry, and every threshold behind every signal traces to a published methodology. See how verification works for the fail-closed gate that decides whether a narrative publishes at all.

Open the product demo How verification works